Selection algorithm based on Kullback-Leibler distance is one of the simplest, fastest, and most effective methods suitable for feature selection of real applications like leak detection systems. However, this method has problems when the training dataset is not large enough. This paper proposes a crossing level value that evaluates the level of overlap between the conditional probability space and the degree of dispersion of each probability to choose the best features before classifying. The evaluation results indicate the proposed method is more stable, more reliable, and has a higher accuracy than the Kullback-Leibler method.